Research On Image Restoration Based On Ensembling Priors Constraint Model

With the rapid development of the social network,IoT and Internet,the digital media technique presents anywhere and has been widely applied on many areas.Image is a significant carrier of the digital media,whose quality plays an important role on the sensing,communication and understanding.During the process of collection,transmission,storage,image will be contaminated with noise,blurring and distortion due to many factors,such as devices,transmission environment and human operation fault.The obtained image is degraded consequently,which has strong influence on image quality and also brings challenging problems to the subsequent recognition,classification and understanding tasks.So,it is a hot topic all the time that aiming to know how to recovery the realistic image from the its degraded version,which has important valuable on both of theory and practice.It is due to that the significant information is lost during the degradation,image restoration can be seen as an ill-posed mathematical problem,which can be changed into a properly posed problem with the help of prior regularization and obtain the approximate solutions.In this paper,we focus on the sparsity and low-rank priors,and develop an in-depth study on three typical image restoration problems that are denoising,deblurring and compressed reconstruction respectively.Also,we mainly focus on the key techniques,including sparse representation,low-rank constraint,nonlocal prior and regularization et al.By embedding the nonlocal prior into the sparsity and low-rank constraint,we design the concrete embedding pattern according to restoration task.Then,we consider to combine them into the same objective function and construct the restoration model with ensemble prior constraint.What we want is to introduce sufficient prior constraint into the restoration model as far as possible,which can meet the requirements of multiple constraints simultaneously and solve some key problems,such as clear content recovery,texture preservation and details reconstruction et al,to further improve the restoration quality.The innovation research results include the following four aspects:1.For the gaussian noise removing from images,the conventional algorithms consider only one of the local or global modeling in the priors,which can not character the structure information of images completely.To capture the latent global information and local sparsity,this paper proposes an image denoising method by ensembling hidden variable and group sparsity.In the proposed method,the hidden vairable is used to character the global information and two priors are proposed to constrain the hidden variable and its high-frequency component,which can solve the separation on low and high frequency components.Meanwhile,a method on patches group construction and its group constraint model are proposed to capture the local sparsity in images.At last,by ensembling the above two priors,the proposed image denoising framework is established.The experimental results on benchmark testing images demonstrates that our proposed method outperforms the current denoising methods.2.For the mixed noising removing from images,on one side,the existing methods adopts the additive model to character the degradation generally,which brings some difficulties to the modeling on inverse problem with two kinds of noise distribution.On the other side,they ignore the effect of nonlocal similarity on structure preservation.Hence,inspired by the degradation model in deblurring problem,different from the conventional additive mixed noise model,the paper proposes a mixed noise removing method with both of nonlocal low-rank and total variation prior constraints.Firstly,the proposed method takes the impulse noise as the blurring kernel and construct the linear degradation model with mixed noise.Secondly,considering the correlation among nonlocal patches,the paper design two kinds of prior constraints by embedding the nonlocal constraint and combine them into the same restoration framework that can be used to character the similarity and smooth on the nonlocal patches.The experimental results under groups of impulse and gaussian noise with different intensity show the priority of our proposed method on mixed noise removing compared to the existing methods.3.The coding accuracy is a significant factor on image deblurring alogrithm based on sparse representation.Nevertheless,the contributions of nonlocal patches on coding structure similarity are ignored within the existing sparse representation-based deblurring methods,which leads to a terrible coding accuracy.To address the problem,inspired by the nonlocal technique,the paper proposes an image deblurring algorithm based on sparsity constraint on patches correlations.Different from the conventional Euclidean measurement,the proposed algorithm explores the structure similarity with sparse self-representation and take the similarity measurement as the weights to constrain the approximation among the sparse coding,which is proved to help to improve the coding accuracy and restoration quality.The experimental results under both of gaussian and union blurring kernels demonstrate that the proposed algorithm can obtain better edge and texture recovery performance than the conventional methods.4.For the image compressed reconstruction problem,it is due to that the reconstruction image can not be obtained firstly,the reconstruction model fails in exploiting the existed nonlocal similarity,which degrades the reconstruction performance on texture or edge preservation.To address the problem,the paper adopts an iterative reconstruction scheme,which explore the nonlocal structures with the reconstruction result in the last iteration and proposes an image compressed reconstruction algorithm.On one side,the proposed algorithm enhances the adaptation of kernel norm by correcting the singular weights in the nonlocal low-rank constraint.On the other side,it designs a 3D sparisty constraint by utilizing the cubic Surfacelet transformation.Then,we construct the undersampling image reconstruction model by putting the above two constraints into the same framework and testing its performance on both of medical images denoising and compressed reconstruction.The experimental results show the advantage of our proposed algorithm on reconstruction quality and details reconstruction with both of visual effect and objective evaluation.